Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

HEC-GCN: Hypergraph Enhanced Cascading Graph Convolution Network for Multi-Behavior Recommendation

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read HEC-GCN claims consistently superior multi-behavior recommendation by combining behavior-specific hypergraphs with cascading graph convolution and contrastive consistency alignment, reporting relative HR@10 gains of 19.20%, 37.45%, and…

desk verdict A well-ablated new architecture for multi-behavior recommendation; the large claimed gains are plausible but not yet backed by uncertainty quantification. read the letter →

arxiv 2412.14476 v1 pith:RVXMTANR submitted 2024-12-19 cs.IR

Yabo Yin , Xiaofei Zhu , Wenshan Wang , Yihao Zhang , Pengfei Wang , Yixing Fan , Jiafeng Guo This is my paper · ORCID
classification cs.IR
keywords multi-behaviorrecommendationhypergraphgraphconvolutionnetworkcontrastivelearningself-supervisedcascadingbehaviordatasparsitycollaborativefiltering
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Multi-behavior recommendation uses auxiliary signals such as views and cart adds to predict the target buy behavior, but existing models focus on fine-grained pairwise interactions and on dependencies between two consecutive behaviors only. This paper argues that both the sparse fine-grained view and the cross-behavior consistency deserve equal treatment. The proposed HEC-GCN therefore pairs, for each behavior, a user-item interaction graph with a learnable hypergraph that captures coarse-grained correlations, runs these in a cascaded manner, and aligns representations through intra- and inter-behavior contrastive losses. The paper reports that this consistently outperforms previous state-of-the-art methods on three public benchmarks, with relative HR@10 improvements of 19.20%, 37.45%, and 13.43% over the best baseline. If correct, the result shows that coarse-grained hypergraph views plus global consistency constraints are a practical way to fight sparsity in multi-behavior recommenders.

What carries the argument

For each behavior $k$, embeddings are learned twice: once by LightGCN-style propagation on the behavior-specific user-item interaction graph, and once by a simplified hypergraph convolution whose hyperedges are produced by a low-rank linear map from the graph embeddings. The two views are integrated with a residual connection from the previous behavior, giving the cascade; a stop-gradient keeps information flowing one way from graph to hypergraph. Two InfoNCE-style losses align these views: an intra-behavior loss matches the interaction-graph embedding with the hypergraph embedding of the same node, and an inter-behavior loss aligns each behavior's embeddings with embeddings from a global graph containing all behaviors. A behavior mutual enhancement module then re-weights behavior embeddings by softmax attention, and multi-task BPR losses with tuned weights combine prediction, contrastive, and regularization terms.

What would settle it

An independent replication that runs HEC-GCN, BCIPM, PKEF, and MB-HGCN under the reported hyperparameter ranges with multiple random seeds would settle the claim: if the relative HR@10 margins on Taobao or Beibei shrink to within run-to-run variance, the consistency claim fails. A cheaper check is the paper's own reordering test: on Taobao, any behavior order other than [view, cart, buy] should degrade performance, and the reverse order should fall below the no-cascade variant, as reported.

Watch

Extended reading notes

Core claim

The central claim is that adding a behavior-specific hypergraph to a cascading graph convolution network, together with a behavior consistency-guided contrastive learning module, yields a recommendation model that is consistently superior to previous state-of-the-art methods on the Beibei, Taobao, and Tmall benchmarks. The paper reports relative HR@10 gains of 19.20%, 37.45%, and 13.43% over the strongest baseline on those datasets, and similar gains in NDCG@10. The architecture attributes the improvement to the hypergraph's ability to capture coarse-grained high-order correlations that complement the sparse fine-grained interaction graph, and to the alignment of embeddings across behaviors through a shared global graph.

Load-bearing premise

The reported superiority assumes the baselines were tuned to comparable strength and that a single-run, leave-one-out evaluation with undisclosed hyperparameter values is stable, and the model also assumes the fixed behavior order (view, cart, buy) is semantically appropriate, which the paper's own reordering experiments show is critical.

Editorial extensions

If this is right

  • Auxiliary behaviors carry substantial predictive signal, with the paper showing that removing the 'view' behavior hurts all models most and that HEC-GCN extracts the most benefit from auxiliary data.
  • The fixed order of behaviors matters for the cascading design, and model designers should place noisier low-intent behaviors earlier in the chain.
  • Contrastive alignment through a global graph is a transferable way to share information across behaviors without forcing behavior-specific representations to be identical.
  • On the target behavior alone, the multi-behavior architecture is less competitive than LightGCN, so the reported gains come specifically from the multi-behavior machinery rather than from the base encoder.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The low-rank parameterized hyperedges effectively learn a soft clustering of users and items per behavior; a natural test is whether the optimal hyperedge count $S$ aligns with the number of interpretable item categories in each dataset.
  • Because the paper reports no seed variance, the large margins on Taobao (37.45%) may partly reflect baseline tuning rather than model advantage; a multi-seed comparison with released code would clarify this.
  • An order-adaptive variant that learns the behavior chain rather than assuming view to cart to buy could extend the method to platforms where behavior semantics differ, though the paper's reordering results suggest the assumption is currently load-bearing.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes HEC-GCN, a multi-behavior recommendation model that combines a global graph encoder, per-behavior LightGCN encoders, per-behavior learnable hypergraphs, cascading propagation from auxiliary behaviors to the target behavior, and two contrastive consistency losses (inter-behavior alignment to the global graph and intra-behavior alignment between graph and hypergraph views). The model is optimized with a multi-task BPR loss plus weighted contrastive terms. On the Beibei, Taobao, and Tmall datasets, Table 2 reports HR@10/NDCG@10 improvements over the best baseline of 19.20%/15.98%, 37.45%/31.47%, and 13.43%/14.82%, respectively. Ablation and sensitivity analyses are used to attribute the gains to the hypergraph, cascading architecture, contrastive losses, and behavior mutual enhancement.

Significance. If the reported empirical gains are stable, HEC-GCN would be a strong new state of the art for multi-behavior recommendation, combining an established cascade paradigm with hypergraph views and contrastive alignment in a coherent, well-motivated architecture. Concrete strengths include a released code repository, ablations for every major component, and purpose-built analyses of interaction density, cold-start robustness, auxiliary-behavior impact, and behavior-order sensitivity. However, the evidence for the headline claim is currently incomplete: the comparisons report no variance or significance information, the final per-dataset hyperparameters are not given, and the target-behavior-only experiment in Section 5.7 shows the model underperforming LightGCN. The central claim is therefore plausible but not yet established to the standard required by the abstract's wording.

major comments (3)
  1. [Section 5.2, Table 2] The abstract states that HEC-GCN is 'consistently superior to previous state-of-the-art methods', but the Table 2 comparison rests on scores for which no variance, confidence interval, or significance test is reported. Section 5.1.3 lists only search ranges for lambda_1, lambda_2, lambda_3, alpha, tau, S, L, and d, and the final values are never stated. Because Figures 6 and 7 show HR@10 varying nontrivially as tau and S change, the reported 19.20%-37.45% relative margins could in part be tuning or seed artifacts. Please provide the final hyperparameters for each dataset, means and standard deviations over multiple runs, and a significance test (e.g., paired bootstrap or Wilcoxon signed-rank) for the Table 2 comparisons; also clarify whether baseline numbers are re-computed under the same split and protocol or taken from prior publications.
  2. [Section 5.7, Figure 4] The manuscript concedes that when only the target behavior is used, HEC-GCN, like the other multi-behavior models, performs below LightGCN. This is inconsistent with the unqualified statement in the abstract that HEC-GCN is 'consistently superior to previous state-of-the-art methods'. The superiority claim should be explicitly restricted to multi-behavior settings in which auxiliary behaviors are available, and the paper should discuss why HEC-GCN cannot gracefully degrade to a competitive single-behavior recommender when auxiliary signals are absent.
  3. [Section 5.8, Figure 5] The cascade component is sharply sensitive to behavior order: the reverse chain [buy -> cart -> view] performs worse than the non-cascading variant, so the cascade can degrade performance when the ordering assumption is wrong. The paper treats this only as support for the natural [view -> cart -> buy] order, but it is also a limitation of the method. Please state whether the same chain order is used for all cascading baselines (e.g., MB-CGCN and CRGCN), and discuss the ordering assumption as a required input and possible source of bias in the Table 2 comparisons.
minor comments (6)
  1. [Throughout] There are numerous typos and misspellings, including 'behanvior' (Section 4.2 title), 'Inter-Bahavior' and 'Intra-Bahavior' (Section 4.3 and Figure 1), 'casual behaviors' for 'causal behaviors' (Section 5.8), and 'Moveover' (Section 2). A careful proofread is needed.
  2. [Section 4.2.2, Eqs. (10)-(13)] The hypergraph incidence is generated as H = E_b W_h and the hypergraph view is E_h = (H H^T) E_b, so the hypergraph embeddings are a deterministic transformation of the same E_b used by the interaction graph. Please clarify what distinguishes the 'coarse-grained' information from a learned low-rank re-encoding of the same node embeddings, and consider reporting an analysis of the learned hyperedges to show that they carry interpretable group structure.
  3. [Section 5.9, Figure 7] The text says 'the optimal performance is achieved by setting a relatively small value of S', but Figure 7 shows clearly lower performance at S=16 than at interior values. Please report the selected S for each dataset and define what 'small' means relative to the number of users or items.
  4. [Section 5.1.4] The leave-one-out protocol should state the ranking candidate set: whether the 10 items for HR@10/NDCG@10 are selected from all items or from a fixed set of sampled negatives, and whether validation is used for early stopping. These details affect comparability with the baselines.
  5. [Equation (26), Section 5.1.3] Please state the value of beta (the L2 regularization coefficient) and whether it is tuned; Section 5.1.3 says the 'regularization coefficient is set to 1e-3' but does not connect this value to beta in Equation (26).
  6. [Tables 3 and 4] Please clarify whether the ablation variants are trained with exactly the same hyperparameters, validation procedure, and random seed as the full model; otherwise the component contributions could be confounded by re-tuning.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported superiority is an empirical benchmark result, not baked into the model's definitions or self-citations.

full rationale

HEC-GCN is evaluated by leave-one-out ranking on held-out last interactions (Section 5.1.4), and the HR@10 and NDCG@10 numbers in Table 2 are measured outputs, not quantities defined by the model's equations. The global graph, cascading LightGCN, parameterized hypergraph (Eqs. 10-13), contrastive losses (Eqs. 16-20), and BPR objective (Eq. 25) are training and architecture components; none is fitted to the test labels, and none of the reported improvements is a renamed fitted parameter. The intra-behavior contrastive loss does align e^{b,k} with e^{h,k}, where e^{h,k} is a linear transform of e^{b,k} (Eqs. 12-13 vs. Eq. 19), but this is an internal regularizer that cannot by itself produce the benchmark scores; the evaluation signal comes from held-out target-behavior interactions. The paper contains no self-citations to the present authors and invokes no uniqueness theorem; the cited prior work supplies standard external components (LightGCN, BPR, InfoNCE, hypergraph convolution). The paper's own Section 5.7 shows HEC-GCN underperforms LightGCN in the target-only regime, which is inconsistent with a circular construction that would always win. Missing variance estimates and undisclosed final hyperparameter values are reproducibility concerns, not circularity, because nothing in the reported pipeline presupposes the test-set outcome.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The model itself is a standard neural recommender with many learned parameters; the items listed are the specifically tuned hyperparameters and unverified structural assumptions that the reported benchmark numbers depend on. The most consequential omission is that final per-dataset values for nearly every tunable coefficient are not disclosed.

free parameters (6)
  • lambda_1, lambda_2, lambda_3 (consistency loss coefficients) = not reported; grid {0, 0.5, 1.0, 1.5, 2.0, 2.5} with sum constraint 3
    These weights control the contribution of inter-behavior graph/hypergraph and intra-behavior losses; chosen per dataset by validation but exact values omitted.
  • alpha (overall consistency loss coefficient) = not reported; grid {0.1, 0.5}
    Scales the entire contrastive loss in Eq. (26); final value not stated.
  • tau (temperature coefficient) = 0.1 (inferred from Fig. 6; not listed in experimental settings)
    Sharpness of the contrastive similarity; Fig. 6 shows a peak at 0.1 but the main experiment section does not state it.
  • S (number of hyperedges per behavior) = not reported; grid {16, 32, 64, 128, 256}, described as 'relatively small'
    Capacity of the hypergraph; final value per dataset is not reported.
  • L (GCN layers per behavior) = not reported; grid {1, 2, 3}
    Depth of LightGCN propagation; final value omitted.
  • Embedding size d = not reported; grid {64, 128}
    Dimension of user/item embeddings; final choice omitted.
assumptions (4)
  • domain assumption The leave-one-out evaluation protocol with BPR pairwise loss and random negative sampling accurately reflects recommendation quality.
    Section 5.1.4 adopts leave-one-out without analyzing its variance or its agreement with online metrics; the strong claims rest on this protocol.
  • domain assumption The behavior order (e.g., view, cart, buy) is known and fixed a priori and is used to structure the cascade.
    Eqs. (6)-(7) initialize each behavior from its predecessor; Section 5.8 shows performance drops under reordering, so the method is sensitive to this assumption.
  • ad hoc to paper The hyperedge matrix H = E_b * W_h, with S learnable hyperedges, can represent the coarse-grained correlations that complement fine-grained interaction data.
    Eqs. (10)-(13) define hyperedges as linear projections of the same behavior embeddings; the paper asserts but does not demonstrate that this captures category-level or interest-group structure.
  • domain assumption The processed versions of Beibei, Taobao and Tmall datasets match prior work's preprocessing, and interaction data are accurate implicit feedback.
    Section 5.1.1 follows previous works [2,35,18] without providing the preprocessing code or verifying label quality.
invented entities (1)
  • Learnable behavior-specific hyperedges (H_U^k, H_I^k)
    purpose: Model coarse-grained, high-order correlations among users and items within each behavior (Eqs. 10-13).
    The hyperedges are low-rank latent factors trained with the recommendation and contrastive losses; no analysis links a hyperedge to a concrete semantic group, so the claim that each 'reflects a specific semantic dependency' (Section 5.9) is unverified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HEC-GCN: Hypergraph Enhanced Cascading Graph Convolution Network for Multi-Behavior Recommendation." pith.science (2026). https://pith.science/paper/RVXMTANR

@misc{pith2026241214476,
  author       = {Pith},
  title        = {Pith review of: HEC-GCN: Hypergraph Enhanced Cascading Graph Convolution Network for Multi-Behavior Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RVXMTANR}},
  note         = {Machine review of arXiv:2412.14476}
}
read the original abstract

Multi-behavior recommendation (MBR) has garnered growing attention recently due to its ability to mitigate the sparsity issue by inferring user preferences from various auxiliary behaviors to improve predictions for the target behavior. Although existing research on MBR has yielded impressive results, they still face two major limitations. First, previous methods mainly focus on modeling fine-grained interaction information between users and items under each behavior, which may suffer from sparsity issue. Second, existing models usually concentrate on exploiting dependencies between two consecutive behaviors, leaving intra- and inter-behavior consistency largely unexplored. To the end, we propose a novel approach named Hypergraph Enhanced Cascading Graph Convolution Network for multi-behavior recommendation (HEC-GCN). To be specific, we first explore both fine- and coarse-grained correlations among users or items of each behavior by simultaneously modeling the behavior-specific interaction graph and its corresponding hypergraph in a cascaded manner. Then, we propose a behavior consistency-guided alignment strategy that ensures consistent representations between the interaction graph and its associated hypergraph for each behavior, while also maintaining representation consistency across different behaviors. Extensive experiments and analyses on three public benchmark datasets demonstrate that our proposed approach is consistently superior to previous state-of-the-art methods due to its capability to effectively attenuate the sparsity issue as well as preserve both intra- and inter-behavior consistencies. The code is available at https://github.com/marqu22/HEC-GCN.git.

Figures

Figures reproduced from arXiv: 2412.14476 by the authors.

Figure 1
Figure 1. The overall architecture of HEC-GCN. We utilize three behaviors (i.e., view, cart and buy) as an example, wherein buy is the target behavior. where 𝐿 is the total number of graph convolution layers. 4.2. Hypergraph Enhanced Cascading Graph Convolution Network We propose a hypergraph enhanced cascading graph convolution network to learn user and item representations from the user-item interaction graph and its corres… view at source ↗
Figure 2
Figure 2. Model performance with respect to different interac￾tion density degrees. • 𝑤∕𝑜 𝐶𝐿𝑎𝑙𝑙: Both intra- and inter-behavior contrastive losses are simultaneously removed. The results are exhibited in [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. The impact of different cold-start ratios on model performance. [view,cart,buy] [cart,buy] [buy] 0.00 0.02 0.04 0.06 0.08 0.10 0.12 NDCG@10 LightGCN Ours BCIPM PKEF MB-HGCN [view,cart,buy] [cart,buy] [buy] 0.00 0.04 0.08 0.12 0.16 0.20 HR@10 LightGCN Ours BCIPM PKEF MB-HGCN [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The Impact of auxiliary behaviors on performance. 12 interactions). We then report the performance of our proposed model for each category in [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Impact of behavior sequence order on performance in the cascading architecture. Yin et al.: Preprint submitted to Elsevier Page 10 of 12 [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: The Impact of temperature coefficient 𝜏. 16 32 64 128 256 Hyperedge number S 0.90 0.92 0.94 0.96 0.98 1.00 Relative NDCG@10 Beibei Taobao Tmall 16 32 64 128 256 Hyperedge number S 0.88 0.90 0.92 0.94 0.96 0.98 1.00 Relative HR@10 Beibei Taobao Tmall [PITH_FULL_IMAGE:f…
Figure 7
Figure 7. Figure 7: The Impact of hyperedge number 𝑆. reason is that casual behaviors, such as “view”, often contain more noise. By positioning these casual behaviors earlier in the sequence, the cascading architecture can progressively refine the learning of user preferences, allowing a …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. MCLMR: A Model-Agnostic Causal Learning Framework for Multi-Behavior Recommendation

    cs.IR 2026-03 conditional novelty 6.0 of 10

    A model-agnostic causal plug-in improves multi-behavior recommenders via backdoor adjustment on user/item bias proxies, MoE aggregation of auxiliaries, and bias-aware contrastive alignment.

Reference graph

Works this paper leans on

41 extracted references · 38 canonical work pages · cited by 1 Pith paper

  1. [1]

    Effi- cientheterogeneouscollaborativefilteringwithoutnegativesampling for recommendation, in: Proceedings of the Thirty-Fourth AAAI Conference on Artificial Intelligence, pp

    Chen, C., Zhang, M., Zhang, Y., Ma, W., Liu, Y., Ma, S., 2020. Effi- cientheterogeneouscollaborativefilteringwithoutnegativesampling for recommendation, in: Proceedings of the Thirty-Fourth AAAI Conference on Artificial Intelligence, pp. 19–26

    work page 2020

  2. [2]

    Multi-behavior recommendation with cascading graph convolution networks, in: WWW, pp

    Cheng, Z., Han, S., Liu, F., Zhu, L., Gao, Z., Peng, Y., 2023. Multi-behavior recommendation with cascading graph convolution networks, in: WWW, pp. 1181–1189

    work page 2023

  3. [3]

    Ding, J., Yu, G., He, X., Quan, Y., Li, Y., Chua, T., Jin, D., Yu, J.,

    work page

  4. [4]

    Fan, H., Zhang, F., Wei, Y., Li, Z., Zou, C., Gao, Y., Dai, Q.,

    work page

  5. [5]

    Gao,C.,He,X.,Gan,D.,Chen,X.,Feng,F.,Li,Y.,Chua,T.,Jin,D.,

    work page

  6. [6]

    Learning to recommend with multiple cascading behaviors

    Gao, C., He, X., Gan, D., Chen, X., Feng, F., Li, Y., Chua, T., Yao, L., Song, Y., Jin, D., 2021. Learning to recommend with multiple cascading behaviors. IEEE Trans. Knowl. Data Eng. 33, 2588–2601

    work page 2021

  7. [7]

    Understanding the difficulty of training deep feedforward neural networks, in: AISTATS, pp

    Glorot, X., Bengio, Y., 2010. Understanding the difficulty of training deep feedforward neural networks, in: AISTATS, pp. 249–256

    work page 2010

  8. [8]

    Self-supervised graph neural networks for multi-behavior recommendation, in: IJCAI, pp

    Gu, S., Wang, X., Shi, C., Xiao, D., 2022. Self-supervised graph neural networks for multi-behavior recommendation, in: IJCAI, pp. 2052–2058

    work page 2022

Show all 41 references
  1. [9]

    Resolving data sparsity by multi-type auxiliary implicit feedback for recom- mender systems

    Guo,G.,Qiu,H.,Tan,Z.,Liu,Y.,Ma,J.,Wang,X.,2017. Resolving data sparsity by multi-type auxiliary implicit feedback for recom- mender systems. Knowl. Based Syst. 138, 202–207

    2017

  2. [10]

    Guo, L., Hua, L., Jia, R., Zhao, B., Wang, X., Cui, B., 2019. Buying or browsing?: Predicting real-time purchasing intent using attention- based deep network with multiple behavior, in: Proceedings of the 25thACMSIGKDDInternationalConferenceonKnowledgeDiscov- ery & Data Mining...

    2019

  3. [11]

    Compressed interaction graph based framework for multi-behavior recommendation, in: Pro- ceedings of the ACM Web Conference 2023, pp

    Guo, W., Meng, C., Yuan, E., He, Z., Guo, H., Zhang, Y., Chen, B., Hu, Y., Tang, R., Li, X., Zhang, R., 2023. Compressed interaction graph based framework for multi-behavior recommendation, in: Pro- ceedings of the ACM Web Conference 2023, pp. 960–970

    2023

  4. [12]

    Lightgcn: Simplifying and powering graph convolution network for recommendation, in: SIGIR, pp

    He, X., Deng, K., Wang, X., Li, Y., Zhang, Y., Wang, M., 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation, in: SIGIR, pp. 639–648

    2020

  5. [13]

    Neural collaborative filtering, in: WWW, pp

    He, X., Liao, L., Zhang, H., Nie, L., Hu, X., Chua, T., 2017. Neural collaborative filtering, in: WWW, pp. 173–182

    2017

  6. [14]

    Recentadvancesinheterogeneousrelationlearning for recommendation, in: Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, pp

    Huang,C.,2021. Recentadvancesinheterogeneousrelationlearning for recommendation, in: Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, pp. 4442–4449

    2021

  7. [15]

    Multi-behavior recommendation with graph convolutional networks, in: SIGIR, pp

    Jin, B., Gao, C., He, X., Jin, D., Li, Y., 2020. Multi-behavior recommendation with graph convolutional networks, in: SIGIR, pp. 659–668

    2020

  8. [16]

    Adam:Amethodforstochasticoptimiza- tion, in: ICLR

    Kingma,D.P.,Ba,J.,2015. Adam:Amethodforstochasticoptimiza- tion, in: ICLR

    2015

  9. [17]

    Bayesian personalized ranking with multi-channel user feedback, in: RecSys, pp

    Loni, B., Pagano, R., Larson, M.A., Hanjalic, A., 2016. Bayesian personalized ranking with multi-channel user feedback, in: RecSys, pp. 361–364

    2016

  10. [18]

    Meng, C., Zhai, C., Yang, Y., Zhang, H., Li, X., 2023a. Par- allel knowledge enhancement based framework for multi-behavior recommendation, in: Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pp. 1797– 1806

  11. [19]

    Hierarchicalprojectionenhanced multi-behavior recommendation, in: Proceedings of the 29th ACM SIGKDDConferenceonKnowledgeDiscoveryandDataMining,pp

    Meng,C.,Zhang,H.,Guo,W.,Guo,H.,Liu,H.,Zhang,Y.,Zheng,H., Tang,R.,Li,X.,Zhang,R.,2023b. Hierarchicalprojectionenhanced multi-behavior recommendation, in: Proceedings of the 29th ACM SIGKDDConferenceonKnowledgeDiscoveryandDataMining,pp. 4649–4660

  12. [20]

    Meng, C., Zhao, Z., Guo, W., Zhang, Y., Wu, H., Gao, C., Li, D., Li, X.,Tang,R.,2024.Coarse-to-fineknowledge-enhancedmulti-interest learningframeworkformulti-behaviorrecommendation.ACMTrans. Inf. Syst. 42, 30:1–30:27

    2024

  13. [21]

    Scattering gcn: Overcoming oversmoothnessingraphconvolutionalnetworks

    Min, Y., Wenkel, F., Wolf, G., 2020. Scattering gcn: Overcoming oversmoothnessingraphconvolutionalnetworks. Advancesinneural information processing systems 33, 14498–14508

    2020

  14. [22]

    Representation learning with contrastive predictive coding

    van den Oord, A., Li, Y., Vinyals, O., 2018. Representation learning with contrastive predictive coding. CoRR abs/1807.03748

    2018 arXiv

  15. [23]

    BPR:bayesianpersonalizedrankingfromimplicitfeedback,in:UAI, pp

    Rendle,S.,Freudenthaler,C.,Gantner,Z.,Schmidt-Thieme,L.,2009. BPR:bayesianpersonalizedrankingfromimplicitfeedback,in:UAI, pp. 452–461

    2009

  16. [24]

    Modelingrelationaldatawithgraphconvolutional networks, in: ESWC, pp

    Schlichtkrull,M.S.,Kipf,T.N.,Bloem,P.,vandenBerg,R.,Titov,I., Welling,M.,2018. Modelingrelationaldatawithgraphconvolutional networks, in: ESWC, pp. 593–607

    2018

  17. [25]

    Relational learning via collective matrix factorization, in: Proceedings of the 14th ACM SIGKDD InternationalConferenceonKnowledgeDiscoveryandDataMining, pp

    Singh, A.P., Gordon, G.J., 2008. Relational learning via collective matrix factorization, in: Proceedings of the 14th ACM SIGKDD InternationalConferenceonKnowledgeDiscoveryandDataMining, pp. 650–658

    2008

  18. [26]

    Tang, H., Liu, J., Zhao, M., Gong, X., 2020. Progressive layered extraction (PLE): A novel multi-task learning (MTL) model for personalized recommendations, in: Proceedings of the 14th ACM Conference on Recommender Systems, pp. 269–278

    2020

  19. [27]

    An empirical study on recommendation with multiple types of feedback, in: SIGKDD, pp

    Tang, L., Long, B., Chen, B., Agarwal, D., 2016. An empirical study on recommendation with multiple types of feedback, in: SIGKDD, pp. 283–292

    2016

  20. [28]

    Hgrec: Group recommendation with hypergraph convolutional networks

    Wang, N., Liu, D., Zeng, J., Mu, L., Li, J., 2024. Hgrec: Group recommendation with hypergraph convolutional networks. IEEE Trans. Comput. Soc. Syst. 11, 4214–4225

    2024

  21. [29]

    ARGO: modelingheterogeneityine-commercerecommendation,in:Interna- tional Joint Conference on Neural Networks, pp

    Wu, D., Luo, X., Ma, Z., Chen, C., Deng, M., Ma, J., 2021. ARGO: modelingheterogeneityine-commercerecommendation,in:Interna- tional Joint Conference on Neural Networks, pp. 1–8

    2021

  22. [30]

    Multi- behaviorenhancedrecommendationwithcross-interactioncollabora- tiverelationmodeling,in:Proceedingsofthe37thIEEEInternational Conference on Data Engineering, pp

    Xia, L., Huang, C., Xu, Y., Dai, P., Lu, M., Bo, L., 2021a. Multi- behaviorenhancedrecommendationwithcross-interactioncollabora- tiverelationmodeling,in:Proceedingsofthe37thIEEEInternational Conference on Data Engineering, pp. 1931–1936

    1931

  23. [31]

    Knowledge-enhanced hierarchical graph transformer network for multi-behavior recommendation, in: Proceedings of the 35th AAAI Conference on Artificial Intelligence, pp

    Xia, L., Huang, C., Xu, Y., Dai, P., Zhang, X., Yang, H., Pei, J., Bo, L., 2021b. Knowledge-enhanced hierarchical graph transformer network for multi-behavior recommendation, in: Proceedings of the 35th AAAI Conference on Artificial Intelligence, pp. 4486–4493

  24. [32]

    Hypergraph contrastive collaborative filtering, in: SIGIR, p

    Xia, L., Huang, C., Xu, Y., Zhao, J., Yin, D., Huang, J., 2022. Hypergraph contrastive collaborative filtering, in: SIGIR, p. 70–79

    2022

  25. [33]

    Cascading residual graph convolutional network for multi-behavior recommendation

    Yan, M., Cheng, Z., Gao, C., Sun, J., Liu, F., Sun, F., Li, H., 2022. Cascading residual graph convolutional network for multi-behavior recommendation. CoRR abs/2205.13128

    2022 arXiv

  26. [34]

    Cascading residual graph convolutional network for multi-behavior recommendation

    Yan, M., Cheng, Z., Gao, C., Sun, J., Liu, F., Sun, F., Li, H., 2023a. Cascading residual graph convolutional network for multi-behavior recommendation. ACM Trans. Inf. Syst. 42

  27. [35]

    MB-HGCN: A hierarchical graph convolutional network for multi-behavior recom- mendation

    Yan, M., Cheng, Z., Sun, J., Sun, F., Peng, Y., 2023b. MB-HGCN: A hierarchical graph convolutional network for multi-behavior recom- mendation. CoRR abs/2306.10679

    arXiv

  28. [36]

    Yan,M.,Liu,F.,Sun,J.,Sun,F.,Cheng,Z.,Han,Y.,2024. Behavior- contextualized item preference modeling for multi-behavior recom- mendation, in: Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 946–955

    2024

  29. [37]

    Improvingusertopic interest profiles by behavior factorization, in: Proceedings of WWW, pp

    Zhao,Z.,Cheng,Z.,Hong,L.,Chi,E.H.,2015. Improvingusertopic interest profiles by behavior factorization, in: Proceedings of WWW, pp. 1406–1416

    2015

  30. [38]

    Towards deeper graph neural networks with differentiable group normaliza- tion

    Zhou,K.,Huang,X.,Li,Y.,Zha,D.,Chen,R.,Hu,X.,2020. Towards deeper graph neural networks with differentiable group normaliza- tion. Advances in neural information processing systems 33, 4917– 4928. Yin et al.:Preprint submitted to Elsevier Page 12 of 12

    2020

  31. [2018]

    3343–3349

    Improving implicit recommender systems with view data, in: IJCAI, pp. 3343–3349

  32. [2019]

    1554–1557

    Neural multi-task recommendation from multi-behavior data, Yin et al.:Preprint submitted to Elsevier Page 11 of 12 in: Proceedings of the 35th IEEE International Conference on Data Engineering, pp. 1554–1557

  33. [2022]

    IEEE Trans

    Heterogeneous hypergraph variational autoencoder for link prediction. IEEE Trans. Pattern Anal. Mach. Intell. 44, 4125–4138

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.